GPT-5.4 Mini Users Report Output Drift

My read: this sounds less like a planned model change and more like either a routing/prompting regression or an integration-layer bug that only shows up when the model is embedded in a tool chain.

• –OpenAI’s own docs frame GPT-5.4 mini as a fallback in ChatGPT and a workhorse for subagents, so even small behavior shifts can ripple through automation stacks.
• –The reported “thoughts” leakage fits a prompt-contamination or output-filtering issue as much as a raw model-quality drop.
• –Home Assistant plus N8N adds another layer of abstraction, which makes it hard to pin the blame on the model alone without controlled A/B tests.
• –If other users can reproduce it across clean API calls, that would be a stronger signal of a real model regression than a single anecdote.
• –For production use, this is a reminder to keep evals, prompt guards, and output sanitizers in place even on “stable” small models.